Research
How Many Exposures Does a World Model Need? Learning Unseen Game Mechanics at Test Time
Researchers introduce MechEval, which labels 69 recurring mechanics across 20 ARC-AGI-3 games and evaluates world-model predictions on the cells each mechanic changes. They also present WARC, a 35M-parameter vision transformer pretrained on other games that updates all parameters after each transition. On held-out games, WARC predicts four in five mechanics with at least 80% accuracy on changed cells, with half reaching that accuracy within about 80 exposures in a single play. In closed-loop play, WARC-Agent raises an explorer's official efficiency score from 0.15 to 0.46, largely through test-time training.
world modelsARC-AGI-3test-time learning